DHE-GD: A Dynamic Hierarchical Ensemble with Group Decision Framework for Credit Card Fraud Diagnosis and Identification

Traditional credit card fraud detection (CCFD) methods have some drawbacks, such as limited capabilities, simplistic rules, and insufficient interpretability in handling highly imbalanced and dynamically evolving fraud patterns. To address these limitations, this paper proposes a Dynamic Hierarchical Ensemble with Group Decision (DHE-GD) framework. This framework significantly enhances detection performance and model robustness through multi-rule collaboration and a dual-objective optimization mechanism. DHE-GD employs three core components: (1) Dynamic hierarchical iteration, recursively refining prediction bias using aggregated results from Borda, Vote, and Naolu rules across layers; (2) Multi-objective optimization, achieving Pareto-optimal weight allocation by minimizing individual JS divergence while maximizing global diversity via convex optimization with theoretical guarantees; (3) Adaptive weight updating, integrating exponential decay mixing coefficients and early stopping for efficient convergence. Experimental results on a German dataset demonstrated that DHE-GD achieves an accuracy of 83.50% and an [Formula: see text]-score of 89.04%, significantly outperforming benchmarks. The model balanced recall (95.04%) and precision (83.75%), minimizing undetected risks while enhancing decision transparency. Optimized via Bayesian hyperparameter tuning, the DHE-GD framework achieved efficient computation with limited computational resources, offering a practical and intelligent solution for CCFD.